A production-grade demand forecasting engine for retail inventory — trained on 5 years of transaction history across 10 stores and 50 SKUs, served through a REST API with sub-second inference.
Five stages turn five years of point-of-sale data into a single number an inventory manager can act on.
5 years of daily sales across 10 stores, 50 items — cleaned and validated.
Lag windows, rolling means, cyclical calendar encodings built per SKU.
Gradient-boosted model (LightGBM) fit across all 500 series jointly.
Time-based holdout — never random split — mirrors real forecasting conditions.
FastAPI endpoint returns a demand prediction with confidence context.
Feature importance shows recent sales momentum — not season or weekday — drives most predictions.
Point-in-time demand prediction, ready to wire into a dashboard, a purchase-order script, or a WhatsApp stock alert.